Search signals, briefings, company results, benchmarks and glossary terms.
Search signals, briefings, company results, benchmarks and glossary terms.
Use this view to inspect the underlying evidence corpus. For ranked developments, decision posture and interpretation, use Signals.
Raw feed or Signals?
Raw feed is chronological evidence. Signals ranks and interprets material change.
Research highlights a 'render confound' in evaluating LLM memory with revising records, where prompt presentation, not underlying mechanism, skews results.
New research introduces Stochastic Reset Pathfinding (SRP), an episodic learning problem for optimizing paths with uncertain edge success probabilities and resets.
Physics-enhanced reinforcement learning (RL) improves sample efficiency for optimal control of complex dynamical systems, addressing a key RL limitation.
Research paper applies Renormalization Group theory to Transformer attention mechanisms, deriving testable predictions for fixed-point geometry and perturbation decay.
Researchers introduced a statistically grounded, threshold-based framework for interpretable classification using Bernoulli Naive Bayes models for medical data.
Research introduces a blueprint for thermodynamic computing leveraging stochastic analog processes to address energy and latency demands of ML workloads.
Research from arXiv explores multi-agent system (MAS) advantages over single-agent systems (SAS) using an information bottleneck perspective.
Research proves a finite-sample formulation gap for physics-informed learning in nonlinear multiscale elliptic equations, improving stability.
Research on physics-based deep learning U-Net for high-resolution precipitation nowcasting (10-90 min period) for urban flood management.
Researchers propose DADiff, a diffusion-driven method for cross-domain policy adaptation in reinforcement learning, addressing dynamics mismatch with limited target domain interaction.
Research demonstrates LLM-driven AutoML using GPT-5, GPT-4o, and Claude Sonnet 4 to autonomously design and refine neural architectures for cross-lingual handwritten OCR.
Research details SpeechGuard, an online defense mechanism to detect and mitigate backdoor attacks targeting speech recognition models at runtime.
Research compares model merging to joint multi-task training for reinforcement learning agents, using Qwen3-8B on AppWorld to test performance.
Research finds Chain-of-Thought (CoT) reasoning in LLMs complicates refusal control, making activation steering less effective than with non-CoT models.
MLLM-DataEngine is a research proposal for a closed-loop system to iteratively generate multimodal instruction tuning data, train, and evaluate MLLMs.
Research explores self-distillation for linear and logistic regression models when original training data is unavailable, using fresh unlabeled covariates.
Research identifies critical vulnerability in facial recognition to intentional electromagnetic interference attacks, extending physical presentation risks.
Research demonstrates fingerprinting federated learning architectures using 5G PHY-layer side channels, even with encrypted payloads.
New research explores performativity in predictive models, where model deployment influences future data, creating feedback loops during retraining.
Research explores using entropy-based features to enhance supervised network traffic classification for anomaly detection, addressing diverse traffic patterns.
Research introduces "prompt echoing" to resolve the "question-first paradox" in Vision-Language Models, improving performance by repeating the question.
New research explores distribution testing against bounded classes of distinguishers to assess if an unknown distribution matches a reference.
Research proposes a privacy-preserving framework using unsupervised keypoints for real-time fall detection, reducing bandwidth needs by compact motion representations.
Research explores embedding inferred behavioral structure into Neural Process-based models for adaptive residential short-term load forecasting.
PagedWeight proposes dynamically quantizing MoE LLM weights to balance memory for model weights and KV cache, improving serving efficiency.
Research finds the Muon optimizer significantly improves agentic reinforcement learning performance (+88% validation success) over AdamW in sparse-reward environments.
Research explores 'Generative Compilation,' using on-the-fly compiler feedback during LLM code generation to improve strict language (e.g., Rust) code quality.
Research proposes a method for generating differentially private synthetic data designed to preserve causal estimands for accurate causal inference.
Research on Backward Conformal Prediction (BCP) proposes a non-conformity score transformation to improve coverage guarantees.
Research provides a statistical interpretation of Evidential Deep Learning (EDL) for uncertainty-aware classification, addressing its theoretical foundations.
Researchers introduced a Graph Neural Network foundation model for event classification, trained on 120 million simulated proton-proton collisions.
Researchers introduced LVSum, a human-annotated benchmark to evaluate multimodal large language models for timestamp-aware long video summarization.
Research proposes a conformal prediction framework for graph-valued outputs, using Z-Gromov-Wasserstein distance for uncertainty quantification.
Research proposes a label-free concept drift detection method for AI models deployed in dynamic environments, enabling MLOps to trigger retraining.
Research identifies 'code-poisoning property inference attacks' as a new threat, where malicious code on hosting platforms could leak private training data properties.
Research demonstrates a large-scale remote sensing VLM achieves strong performance using a simplified architecture, challenging specialized designs.
AuditVotes introduces a framework to improve the certified robustness and accuracy of Graph Neural Networks (GNNs) against adaptive attacks.
Research explores dynamic budget allocation for multi-turn LLM evaluation to efficiently identify jailbreaks or task completion in conversational settings.
DualKV introduces a FlashAttention optimization for RL training with large rollouts and long contexts, reducing compute and memory costs.
Research identifies 'topology overfitting' in GNNs for power grids, where single-task models fail on unseen grid structures despite low in-distribution error.
Research paper proposes PASs-MoE, a method to mitigate router and expert co-drift in LoRA-based Mixture-of-Experts for continual learning in MLLMs.
Research proposes a new variational inference method for evidential deep learning (EDL) to address limitations in quantifying epistemic uncertainty.
Research explores pseudo-calibration with conformal prediction to maintain marginal coverage guarantees under bounded label-conditional covariate shift.
Research proposes a classifier-based adaptive stopping framework for sampling kernels in Bayesian inference to improve MCMC efficiency.
Research introduces Factorized Neural Operators to better capture multiscale physical behavior by decomposing dynamic and persistent responses.
New research proposes an improved method for optimizing orthogonal matrices, potentially scaling to thousands of constraints, building on the Landing algorithm.
Research finds that larger models do not consistently outperform smaller ones under data scarcity, challenging scaling law assumptions.
RhinoVLA is a proposed Vision-Language-Action model addressing real-time deployment challenges on edge hardware by reducing VLM visual and context tokens.
Research highlights the untapped potential of UMAP's internal k-nearest-neighbor graph for high-dimensional data sensemaking, often overlooked in favor of its 2D projection.
Research finds boundary-seeking distillation, effective for classifiers, fails to transfer knowledge effectively for autoencoders in data-free settings.
© 2026 OneBench: AI Insights. All rights reserved.
Evidence before opinion